<![CDATA[The spread of fraudulent content on social media X has become an important issue because perpetrators often use persuasive, urgent, and misleading language to influence users to transfer money, share personal data, or access suspicious links. This research evaluates the performance of IndoBERT for binary classification of fraud and non-fraud Indonesian-language posts on social media X using a two-stage fine-tuning design. The dataset consists of 5,235 manually labeled posts, including 2,557 fraud and 2,678 non-fraud instances. In Stage 1, four IndoBERT variants, namely indobert-base-p1, indobert-base-p2, indobert-large-p1, and indobert-large-p2, were compared using a uniform training configuration to identify the best model. The results showed that indobert-large-p1 at epoch 5 achieved the best performance, with a validation F1-score for the fraud class of 0.8898 and a test accuracy of 0.8989. In Stage 2, the selected model was re-evaluated through a controlled grid search by varying epoch, learning rate, and batch size. Although the best Stage 2 configuration improved the validation F1-score to 0.8975, it did not surpass the best Stage 1 model on the test set. These findings indicate that IndoBERT is effective for fraud detection and that a two-stage evaluation design supports more systematic model selection.]]>
Copyrights © 2026
